transcriber-local

name	transcriber-local
description	Privacy-focused offline transcription using local Whisper models. Use when data privacy is critical, offline operation is needed, or avoiding API costs. Runs entirely on your machine.
version	1.0.0

Transcriber Local

Note: Review PROFILE.md for user-specific model preferences, output formatting, and privacy settings.

Master Briefing: Global brand voice at ~/.superskills/master-briefing.yaml applies automatically. Skill profile overrides when conflicts exist.

Privacy-first offline transcription using local Whisper models. Perfect for sensitive content, offline workflows, and cost-conscious projects.

Tools

LocalTranscriber.py (in src/):

• Runs Whisper models locally (no cloud API)
• Multiple model sizes (tiny, base, small, medium, large)
• GPU acceleration support (CUDA)
• Complete data privacy (nothing leaves your machine)
• No usage limits or API costs
• Word-level timestamps
• Multiple output formats (TXT, JSON, SRT, VTT)
• Batch processing support

Core Workflow

1. Model Selection

• Choose model size based on accuracy/speed trade-off:
  • tiny: Fastest, lowest accuracy (~1GB RAM)
  • base: Fast, decent accuracy (~1GB RAM)
  • small: Balanced (~2GB RAM) [Recommended]
  • medium: High accuracy (~5GB RAM)
  • large: Highest accuracy (~10GB RAM)
• Models are cached after first download

2. Transcription

• Load model into memory (cached for subsequent runs)
• Process audio/video file locally
• Extract word-level timestamps
• Detect language automatically
• Generate confidence scores

3. Delivery

• Export in requested format (TXT, JSON, SRT, VTT)
• Save locally with metadata
• Compatible with same workflow integrations as cloud transcriber

Usage

Basic Local Transcription:

from superskills.transcriber_local.src import LocalTranscriber

transcriber = LocalTranscriber(model_size="small")
result = transcriber.transcribe("recording.mp3")
print(result.transcript)

GPU Acceleration:

transcriber = LocalTranscriber(model_size="medium", device="cuda")
result = transcriber.transcribe("video.mp4")

Batch Processing:

files = ["session1.mp3", "session2.mp3", "session3.mp3"]
results = transcriber.transcribe_batch(files)

Privacy Benefits

✅ No Data Transmission: All processing happens locally ✅ GDPR Compliant: No third-party data sharing ✅ Offline Capable: Works without internet ✅ No Usage Limits: Transcribe unlimited content ✅ Zero API Costs: One-time model download only

Performance Comparison

Model	Speed (1hr audio)	Accuracy	VRAM
tiny	~2 min	Good	1GB
base	~3 min	Better	1GB
small	~5 min	Great	2GB
medium	~12 min	Excellent	5GB
large	~30 min	Best	10GB

Times on CPU. GPU acceleration 3-10x faster.

Integration with Other Skills

→ author: Converts transcripts into structured content → editor: Polishes transcripts for publication → narrator: Creates audio from edited transcripts (round-trip workflow) → coursepackager: Packages transcripts into training materials

Setup

Install Dependencies:

pip install openai-whisper
# Optional GPU support:
pip install openai-whisper[cuda]  # NVIDIA
pip install openai-whisper[metal] # Apple Silicon

First Run:

transcriber = LocalTranscriber(model_size="small")
# Model downloads automatically (~500MB for small model)

When to Use Local vs Cloud

Use transcriber-local when:

• Privacy/compliance requires on-premise processing
• Working offline or with unreliable internet
• Processing large volumes (API costs add up)
• Need complete control over data flow

Use transcriber (cloud) when:

• Need fastest possible results (cloud GPUs)
• Don't have powerful local hardware
• Prefer not managing local models
• Working with very large files (>10GB)

Quality Checklist

Before finalizing transcripts:

• Choose appropriate model size for your accuracy needs
• Verify language detection is correct
• Check timestamps align with audio/video
• Review confidence scores for low-quality sections
• Test with sample before batch processing
• Ensure sufficient disk space for model cache

Troubleshooting

Model Download Fails:

• Check internet connection (first-time only)
• Verify disk space (~500MB-3GB per model)
• Try different model size

Out of Memory:

• Use smaller model (tiny or base)
• Process shorter segments
• Close other applications
• Add more RAM or use cloud transcriber

Slow Performance:

• Use GPU if available
• Choose smaller model
• Process shorter files
• Consider cloud transcriber for speed

Examples

Secure Client Recordings:

# Sensitive coaching sessions - never leave your machine
transcriber = LocalTranscriber(model_size="medium")
result = transcriber.transcribe("client_session.mp3")

Offline Field Work:

# No internet? No problem
transcriber = LocalTranscriber(model_size="base")
results = transcriber.transcribe_batch(["interview1.wav", "interview2.wav"])

Cost-Effective Bulk Processing:

# Process 100 hours of content with zero API costs
transcriber = LocalTranscriber(model_size="small", device="cuda")
for file in large_audio_collection:
    result = transcriber.transcribe(file)

Security & Compliance

• HIPAA: Suitable for healthcare recordings (no PHI transmission)
• GDPR: Compliant (no personal data leaves your control)
• SOC 2: Compatible with on-premise requirements
• Zero Trust: No external API dependencies

Next Steps

After transcription:

1. Review and edit transcripts
2. Extract key quotes or highlights
3. Create structured content with author
4. Package into course materials with coursepackager
5. Generate video captions (SRT/VTT output)